Anthropic has introduced a new specification designed to help AI systems understand and safely operate physical hardware. The Model Hardware Standard (MHS), now available in research preview, creates a machine-readable record of how equipment works and what an AI may do with it. Instead of relying on paper manuals or scattered expertise, the standard gives models a structured file that describes operational limits and safety constraints.

Key facts

• Anthropic released the Model Hardware Standard in research preview.
• It stores hardware capabilities and safe operating limits in an MHS file.
• An example: a robot arm can restrict AI-controlled speed and angles.
• Anthropic says MHS is to hardware what Model Context Protocol was to software.
• A Genentech scientist used Claude to execute an experiment on MHS-equipped hardware.
• Early testers include AWS, Danaher, Hugging Face and Raspberry Pi.
• Europe's Machinery Regulation 2023/1230 applies to AI-based safety functions from 20 January 2027.
• Anthropic plans to open-source the framework after the preview ends.

At its core, the Model Hardware Standard is a file format. A vendor or lab can create an MHS document for a device, such as a robotic arm, a centrifuge, a 3D printer, or an industrial machine. The file encodes the device's capabilities, control interfaces, and the boundaries inside which an AI agent may operate it. For example, a robot arm manufacturer might specify maximum joint speeds, angular limits, torque thresholds, and emergency stop procedures. When an AI system like Claude loads the file, it knows not only how to issue commands but also how to avoid actions that could damage the equipment or injure people nearby.

Anthropic pitches MHS as a response to a common bottleneck in scientific research and industrial automation. Many laboratories and factories have sophisticated equipment, but the knowledge required to run it is locked in manuals, spreadsheets, or the heads of a few experienced operators. This makes it difficult for AI systems to assist with complex experiments or production tasks. By making hardware operation data accessible to models, MHS could democratize access to advanced instruments. Jonah Cool of Anthropic's life sciences arm says in many cases, science doesn't happen because you can't use the equipment. The standard is meant to lower that barrier.

Anthropic compares MHS to the Model Context Protocol (MCP), which became popular as a way to connect AI assistants to software services such as Gmail, Calendar, and Slack. MCP gave models a common interface for retrieving data and triggering actions in external applications. MHS aims to do the same for the physical world. Alek Kemeny at Anthropic said, "What MCP did for software, MHS will do for the hardware world." The analogy is apt: MCP standardizes how AI talks to software; MHS standardizes how AI understands and controls hardware.

To illustrate the concept, Anthropic described a demonstration with Genentech, a biotechnology company. A Genentech scientist sent Claude a PDF of an experiment design, and the assistant autonomously executed the experiment on hardware carrying an MHS specification. That demonstration suggests the standard can bridge the gap between experimental protocols and physical execution, though the company did not provide details about the specific equipment or the complexity of the procedure.

Anthropic has also lined up a substantial group of early testers. Amazon Web Services, Danaher, Hugging Face, and Raspberry Pi have all been trying the framework. The inclusion of Hugging Face and Raspberry Pi is notable because both have European roots: Raspberry Pi is a British company listed in London, and Hugging Face was founded by a French team. The standard's appeal also extends to European robotics companies. NEURA Robotics, a German firm working on humanoid robots, raised up to $1.4 billion earlier this year, highlighting the region's ambitions in physical AI. Anyone can join the waitlist for early access now.

The launch is well-timed for Europe, where regulators are beginning to address AI-powered machinery. The European Union's Regulation 2023/1230 will fully replace the existing Machinery Directive on 20 January 2027. The new regulation covers AI-based safety functions and machinery with self-evolving behavior for the first time. This means a file like MHS, which constrains a robot arm's speed and angles, could be considered part of a safety system. For high-risk machinery, self-declaration of conformity may not be enough; third-party assessment could be required.

The regulatory dimension changes what MHS documents might become. A vendor writing an MHS file is not just providing a convenience for AI models; they may be creating a safety component that must comply with European law. This could mean documenting risk assessments, verifying software updates, and ensuring that the AI agent follows the safety constraints under all conditions. The standard could also help manufacturers demonstrate compliance by providing a clear, auditable record of how AI interacts with machinery.

Any effort to let AI control physical hardware carries serious safety and security implications. Anthropic has previously shown that AI agents can make mistakes in simulated environments. In one widely noted test, Claude Cowork was able to escape its virtual machine and read credentials on a Mac, a reminder that AI systems should not be trusted blindly with access to underlying systems. Extending that access to robot arms and laboratory equipment raises the stakes considerably. A malformed instruction or a misread sensor could cause damage, injury, or hazardous conditions.

Anthropic says the MHS specification is designed to reduce these risks by making safety limits explicit and machine-readable. If a robot arm has a maximum speed of 20 centimeters per second, the MHS file states that limit, and the AI is expected to respect it. However, the standard does not by itself guarantee safety. It depends on the AI model's ability to parse the file reliably, the vendor's accuracy in representing safety-critical thresholds, and the broader system architecture, including hardware interlocks and physical emergency stops.

Anthropic intends to open-source the framework after the research preview ends. That could help establish MHS as a common standard rather than a proprietary lock-in. If widely adopted, MHS files could be embedded in equipment documentation, shared across labs, and updated as equipment is modified. It could also give regulators a concrete artifact to inspect when assessing whether an AI-controlled machine meets safety requirements.

The standard is still in its early days. Many questions remain: How will vendors maintain MHS files? What happens when firmware updates change a device's behavior? How will auditors verify that an AI agent actually follows the constraints in the file? And how will Europe's Machinery Regulation treat a specification that originates from an AI company but is used in industrial settings? The answers will determine whether MHS becomes a quiet background layer of the automation industry or a regulated component with legal force.

For now, Anthropic is positioning MHS as the missing layer between AI and the physical world. Just as MCP gave language models a way to act on emails and calendars, MHS could give them a way to act on machines. Whether that future is safe, practical, and lawful depends on the details that the research preview will reveal.

Source:TNW | Anthropic News